On the Anatomy of Predictive Models for Accelerating GPUConvolution Kernels and Beyond